You're looking at Less Wrong's discussion board. This includes all posts, including those that haven't been promoted to the front page yet. For more information, see About Less Wrong.

shminux comments on Open Thread, May 19 - 25, 2014 - Less Wrong Discussion

2 Post author: somnicule 19 May 2014 04:49AM

You are viewing a comment permalink. View the original post to see all comments and the full post content.

Article Navigation

Comments (289)
Tags: open_thread

Comments (289)

You are viewing a single comment's thread. Show more comments above.

Comment author: shminux 20 May 2014 11:56:31PM *  0 points [-]

Sorry, my writing can get kind of dense.

If you want people to read what you write, learn to write in a readable way.

Looked at your write-up again... Still no summary of what it is about. Something along the lines of (total BS follows, sorry, I have no clue what you are writing about, since the essay is unreadable as is): "This essay outlines the issues with AIXI built on Solomonoff induction and suggests a number of improvements, such as extending algorithmic calculus with interactive calculus. This extension removes hidden infinities inherent in the existing AIXI models and allows <some benefit>."

Comment author: Transfuturist 20 May 2014 11:57:48PM *  1 point [-]

I'm in the process of writing summaries. I replied as soon as I read your response.

If you want people to read what you write, learn to write in a readable way.

You are pretty much the first person to give me feedback on this. I do not have an accurate representation as to how opaque this is at all.

In algorithmic representations:

  1. Separate hypotheses are inseparable.
  2. Hypotheses are required to be a complete world model (account for every part of the input).
  3. Conflicting hypotheses are not able to be held simultaneously. This stems mainly from there being no requirement for running in a finite amount of space and time.

There are other issues with Solomonoff induction in its current form, such as an inability to tolerate error, an inability to separate input in the first place, and an inability to exchange error for simplicity, among others. Some of these are addressable with this particular extension of SI; some are addressable with other extensions.

There is a similar intuition about nondeterministic hypotheses and a requirement that only part of the hypothesis must match the output, as nondeterministic Turing machines can be simulated by deterministic Turing machines via the simulation of every possible execution flow, but that strikes me as somewhat dodgy.

Comment author: Transfuturist 21 May 2014 12:39:09AM *  0 points [-]

How's that? Every few lines, I give a summary of each subsection. I even double-spaced it, in case that was bothering you.

Powered by Reddit Powered by Reddit